Self-Evolving Agents via Likelihood-Guided Tool-Space Optimization

Paper: Self-Evolving Agents via Likelihood-Guided Tool-Space Optimization (v1) https://arxiv.org/html/2609.34151v1

LOTS: keeping a tool menu useful across recurring work

Treating a tool menu as reusable deployment state changes the question from “Which tool fits this request?” to “Which evidence keeps helping this kind of work?” LOTS makes that distinction measurable.

The measurement starts after an answer exists. LOTS freezes the realized output and its recorded supporting context. It removes evidence attributable to one tool and evaluates the same answer again. The contribution is mean output-token log-likelihood under the original context minus the corresponding mean after the intervention. This is a change in support for a fixed answer. It neither generates a replacement answer nor certifies that the original one was correct.

One request gives one observation. Contributions are averaged over usable, exposed requests of the same task type, producing tool priorities. A top-budget subset becomes the stored menu for later requests. That separates measuring a contribution from deciding how many tools to keep. Likelihood scoring uses no correctness labels; the reported experimental budget selection uses fitting labels.

The controlled TGB benchmark contains 4,000 requests, ten task families and fifteen tools. Each family has eighty fitting requests and 320 evaluation requests. For Qwen3-8B, accuracy increases from 56.5% with the full menu to 81.5% with LOTS, a 25-percentage-point difference in that setting. Reported serving-prompt token reductions of 33–46% concern the same controlled registry comparison, rather than universal wall-clock acceleration.

The original animated report example makes the fixed-answer intervention visible. Its practical question is whether the retained evidence keeps supporting recurring work. The resulting menu still needs task evaluation: a model can rely on evidence that supports an incorrect answer.

Watch the companion English animation with external English, Simplified Chinese or bilingual captions.

Paper, exact arXiv v1: https://arxiv.org/html/2609.34151v1

#ToolUse #AIAgents #AIResearch